Re: Auto-tuning work_mem and maintenance_work_mem

Josh Berkus <josh@agliodbs.com> wrote:
> On 10/11/2013 01:11 PM, Bruce Momjian wrote:
>> In summary, I think we need to:
>>
>> *  decide on new defaults for work_mem and maintenance_work_mem
>> *  add an initdb flag to allow users/packagers to set shared_bufffers?
>> *  add an autovacuum_work_mem setting?
>> *  change the default for temp_buffers?
>
> If we're changing defaults, bgwriter_lru_maxpages and vacuum_cost_limit
> could also use a bump; those thresholds were set for servers with < 1GB
> of RAM.

+1 on those.

Also, I have often had to bump cpu_tuple_cost into the 0.03 to 0.05
range to get a good plan.  In general, this makes the exact
settings of *_page_cost less fussy, and I have hit situations where
I was completely unable to get a good plan to emerge without
bumping cpu_tuple_cost relative to the other cpu costs.  I know that
it's possible to engineer a workload that shows any particular cost
adjustment to make things worse, but in real-life production
environments I have never seen an increase in this range make plan
choice worse.

Regarding the settings which have been the center of attention for
most of this thread, I have had very good luck with starting
work_mem at machine RAM * 0.25 / max_connections.  I get the
impression that Josh regards that as too low.  My guess is that he
deals more with data warehouse reporting systems than I do, where
larger settings are both more beneficial and less likely to cause
memory exhaustion than the typical systems I've worked with.  That
is the big problem with auto-configuration -- it depends so much on
the workload.  In the long run, an admission control policy and/or
adaptive configuration based on the observed workload seems like
what we *really* need.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company